linux/fs/notify/notification.c

422 lines
12 KiB
C
Raw Normal View History

fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-21 21:01:20 +00:00
/*
* Copyright (C) 2008 Red Hat, Inc., Eric Paris <eparis@redhat.com>
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2, or (at your option)
* any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; see the file COPYING. If not, write to
* the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
*/
/*
* Basic idea behind the notification queue: An fsnotify group (like inotify)
* sends the userspace notification about events asyncronously some time after
* the event happened. When inotify gets an event it will need to add that
* event to the group notify queue. Since a single event might need to be on
* multiple group's notification queues we can't add the event directly to each
* queue and instead add a small "event_holder" to each queue. This event_holder
* has a pointer back to the original event. Since the majority of events are
* going to end up on one, and only one, notification queue we embed one
* event_holder into each event. This means we have a single allocation instead
* of always needing two. If the embedded event_holder is already in use by
* another group a new event_holder (from fsnotify_event_holder_cachep) will be
* allocated and used.
*/
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-21 21:01:20 +00:00
#include <linux/fs.h>
#include <linux/init.h>
#include <linux/kernel.h>
#include <linux/list.h>
#include <linux/module.h>
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-21 21:01:20 +00:00
#include <linux/mount.h>
#include <linux/mutex.h>
#include <linux/namei.h>
#include <linux/path.h>
#include <linux/slab.h>
#include <linux/spinlock.h>
#include <asm/atomic.h>
#include <linux/fsnotify_backend.h>
#include "fsnotify.h"
static struct kmem_cache *fsnotify_event_cachep;
static struct kmem_cache *fsnotify_event_holder_cachep;
/*
* This is a magic event we send when the q is too full. Since it doesn't
* hold real event information we just keep one system wide and use it any time
* it is needed. It's refcnt is set 1 at kernel init time and will never
* get set to 0 so it will never get 'freed'
*/
static struct fsnotify_event q_overflow_event;
static atomic_t fsnotify_sync_cookie = ATOMIC_INIT(0);
/**
* fsnotify_get_cookie - return a unique cookie for use in synchronizing events.
* Called from fsnotify_move, which is inlined into filesystem modules.
*/
u32 fsnotify_get_cookie(void)
{
return atomic_inc_return(&fsnotify_sync_cookie);
}
EXPORT_SYMBOL_GPL(fsnotify_get_cookie);
/* return true if the notify queue is empty, false otherwise */
bool fsnotify_notify_queue_is_empty(struct fsnotify_group *group)
{
BUG_ON(!mutex_is_locked(&group->notification_mutex));
return list_empty(&group->notification_list) ? true : false;
}
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-21 21:01:20 +00:00
void fsnotify_get_event(struct fsnotify_event *event)
{
atomic_inc(&event->refcnt);
}
void fsnotify_put_event(struct fsnotify_event *event)
{
if (!event)
return;
if (atomic_dec_and_test(&event->refcnt)) {
if (event->data_type == FSNOTIFY_EVENT_PATH)
path_put(&event->path);
BUG_ON(!list_empty(&event->private_data_list));
kfree(event->file_name);
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-21 21:01:20 +00:00
kmem_cache_free(fsnotify_event_cachep, event);
}
}
struct fsnotify_event_holder *fsnotify_alloc_event_holder(void)
{
return kmem_cache_alloc(fsnotify_event_holder_cachep, GFP_KERNEL);
}
void fsnotify_destroy_event_holder(struct fsnotify_event_holder *holder)
{
kmem_cache_free(fsnotify_event_holder_cachep, holder);
}
/*
* Find the private data that the group previously attached to this event when
* the group added the event to the notification queue (fsnotify_add_notify_event)
*/
struct fsnotify_event_private_data *fsnotify_remove_priv_from_event(struct fsnotify_group *group, struct fsnotify_event *event)
{
struct fsnotify_event_private_data *lpriv;
struct fsnotify_event_private_data *priv = NULL;
assert_spin_locked(&event->lock);
list_for_each_entry(lpriv, &event->private_data_list, event_list) {
if (lpriv->group == group) {
priv = lpriv;
list_del(&priv->event_list);
break;
}
}
return priv;
}
/*
* Check if 2 events contain the same information. We do not compare private data
* but at this moment that isn't a problem for any know fsnotify listeners.
*/
static bool event_compare(struct fsnotify_event *old, struct fsnotify_event *new)
{
if ((old->mask == new->mask) &&
(old->to_tell == new->to_tell) &&
(old->data_type == new->data_type) &&
(old->name_len == new->name_len)) {
switch (old->data_type) {
case (FSNOTIFY_EVENT_INODE):
/* remember, after old was put on the wait_q we aren't
* allowed to look at the inode any more, only thing
* left to check was if the file_name is the same */
if (!old->name_len ||
!strcmp(old->file_name, new->file_name))
return true;
break;
case (FSNOTIFY_EVENT_PATH):
if ((old->path.mnt == new->path.mnt) &&
(old->path.dentry == new->path.dentry))
return true;
break;
case (FSNOTIFY_EVENT_NONE):
if (old->mask & FS_Q_OVERFLOW)
return true;
else if (old->mask & FS_IN_IGNORED)
return false;
return false;
};
}
return false;
}
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-21 21:01:20 +00:00
/*
* Add an event to the group notification queue. The group can later pull this
* event off the queue to deal with. If the event is successfully added to the
* group's notification queue, a reference is taken on event.
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-21 21:01:20 +00:00
*/
int fsnotify_add_notify_event(struct fsnotify_group *group, struct fsnotify_event *event,
struct fsnotify_event_private_data *priv)
{
struct fsnotify_event_holder *holder = NULL;
struct list_head *list = &group->notification_list;
struct fsnotify_event_holder *last_holder;
struct fsnotify_event *last_event;
int ret = 0;
/*
* There is one fsnotify_event_holder embedded inside each fsnotify_event.
* Check if we expect to be able to use that holder. If not alloc a new
* holder.
* For the overflow event it's possible that something will use the in
* event holder before we get the lock so we may need to jump back and
* alloc a new holder, this can't happen for most events...
*/
if (!list_empty(&event->holder.event_list)) {
alloc_holder:
holder = fsnotify_alloc_event_holder();
if (!holder)
return -ENOMEM;
}
mutex_lock(&group->notification_mutex);
if (group->q_len >= group->max_events) {
event = &q_overflow_event;
ret = -EOVERFLOW;
/* sorry, no private data on the overflow event */
priv = NULL;
}
spin_lock(&event->lock);
if (list_empty(&event->holder.event_list)) {
if (unlikely(holder))
fsnotify_destroy_event_holder(holder);
holder = &event->holder;
} else if (unlikely(!holder)) {
/* between the time we checked above and got the lock the in
* event holder was used, go back and get a new one */
spin_unlock(&event->lock);
mutex_unlock(&group->notification_mutex);
goto alloc_holder;
}
if (!list_empty(list)) {
last_holder = list_entry(list->prev, struct fsnotify_event_holder, event_list);
last_event = last_holder->event;
if (event_compare(last_event, event)) {
spin_unlock(&event->lock);
mutex_unlock(&group->notification_mutex);
if (holder != &event->holder)
fsnotify_destroy_event_holder(holder);
return -EEXIST;
}
}
group->q_len++;
holder->event = event;
fsnotify_get_event(event);
list_add_tail(&holder->event_list, list);
if (priv)
list_add_tail(&priv->event_list, &event->private_data_list);
spin_unlock(&event->lock);
mutex_unlock(&group->notification_mutex);
wake_up(&group->notification_waitq);
return ret;
}
/*
* Remove and return the first event from the notification list. There is a
* reference held on this event since it was on the list. It is the responsibility
* of the caller to drop this reference.
*/
struct fsnotify_event *fsnotify_remove_notify_event(struct fsnotify_group *group)
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-21 21:01:20 +00:00
{
struct fsnotify_event *event;
struct fsnotify_event_holder *holder;
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-21 21:01:20 +00:00
BUG_ON(!mutex_is_locked(&group->notification_mutex));
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-21 21:01:20 +00:00
holder = list_first_entry(&group->notification_list, struct fsnotify_event_holder, event_list);
event = holder->event;
spin_lock(&event->lock);
holder->event = NULL;
list_del_init(&holder->event_list);
spin_unlock(&event->lock);
/* event == holder means we are referenced through the in event holder */
if (holder != &event->holder)
fsnotify_destroy_event_holder(holder);
group->q_len--;
return event;
}
/*
* This will not remove the event, that must be done with fsnotify_remove_notify_event()
*/
struct fsnotify_event *fsnotify_peek_notify_event(struct fsnotify_group *group)
{
struct fsnotify_event *event;
struct fsnotify_event_holder *holder;
BUG_ON(!mutex_is_locked(&group->notification_mutex));
holder = list_first_entry(&group->notification_list, struct fsnotify_event_holder, event_list);
event = holder->event;
return event;
}
/*
* Called when a group is being torn down to clean up any outstanding
* event notifications.
*/
void fsnotify_flush_notify(struct fsnotify_group *group)
{
struct fsnotify_event *event;
struct fsnotify_event_private_data *priv;
mutex_lock(&group->notification_mutex);
while (!fsnotify_notify_queue_is_empty(group)) {
event = fsnotify_remove_notify_event(group);
/* if they don't implement free_event_priv they better not have attached any */
if (group->ops->free_event_priv) {
spin_lock(&event->lock);
priv = fsnotify_remove_priv_from_event(group, event);
spin_unlock(&event->lock);
if (priv)
group->ops->free_event_priv(priv);
}
fsnotify_put_event(event); /* matches fsnotify_add_notify_event */
}
mutex_unlock(&group->notification_mutex);
}
static void initialize_event(struct fsnotify_event *event)
{
event->holder.event = NULL;
INIT_LIST_HEAD(&event->holder.event_list);
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-21 21:01:20 +00:00
atomic_set(&event->refcnt, 1);
spin_lock_init(&event->lock);
event->path.dentry = NULL;
event->path.mnt = NULL;
event->inode = NULL;
event->data_type = FSNOTIFY_EVENT_NONE;
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-21 21:01:20 +00:00
INIT_LIST_HEAD(&event->private_data_list);
event->to_tell = NULL;
event->file_name = NULL;
event->name_len = 0;
event->sync_cookie = 0;
}
/*
* fsnotify_create_event - Allocate a new event which will be sent to each
* group's handle_event function if the group was interested in this
* particular event.
*
* @to_tell the inode which is supposed to receive the event (sometimes a
* parent of the inode to which the event happened.
* @mask what actually happened.
* @data pointer to the object which was actually affected
* @data_type flag indication if the data is a file, path, inode, nothing...
* @name the filename, if available
*/
struct fsnotify_event *fsnotify_create_event(struct inode *to_tell, __u32 mask, void *data,
int data_type, const char *name, u32 cookie,
gfp_t gfp)
{
struct fsnotify_event *event;
event = kmem_cache_alloc(fsnotify_event_cachep, gfp);
if (!event)
return NULL;
initialize_event(event);
if (name) {
event->file_name = kstrdup(name, gfp);
if (!event->file_name) {
kmem_cache_free(fsnotify_event_cachep, event);
return NULL;
}
event->name_len = strlen(event->file_name);
}
event->sync_cookie = cookie;
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-21 21:01:20 +00:00
event->to_tell = to_tell;
switch (data_type) {
case FSNOTIFY_EVENT_FILE: {
struct file *file = data;
struct path *path = &file->f_path;
event->path.dentry = path->dentry;
event->path.mnt = path->mnt;
path_get(&event->path);
event->data_type = FSNOTIFY_EVENT_PATH;
break;
}
case FSNOTIFY_EVENT_PATH: {
struct path *path = data;
event->path.dentry = path->dentry;
event->path.mnt = path->mnt;
path_get(&event->path);
event->data_type = FSNOTIFY_EVENT_PATH;
break;
}
case FSNOTIFY_EVENT_INODE:
event->inode = data;
event->data_type = FSNOTIFY_EVENT_INODE;
break;
case FSNOTIFY_EVENT_NONE:
event->inode = NULL;
event->path.dentry = NULL;
event->path.mnt = NULL;
break;
default:
BUG();
}
event->mask = mask;
return event;
}
__init int fsnotify_notification_init(void)
{
fsnotify_event_cachep = KMEM_CACHE(fsnotify_event, SLAB_PANIC);
fsnotify_event_holder_cachep = KMEM_CACHE(fsnotify_event_holder, SLAB_PANIC);
initialize_event(&q_overflow_event);
q_overflow_event.mask = FS_Q_OVERFLOW;
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-21 21:01:20 +00:00
return 0;
}
subsys_initcall(fsnotify_notification_init);